Cost control: the cheapest way to get the result

Model routing, caching, off-peak batching, when to move off another provider's subscription, and the red flags that you are overpaying — or using the wrong tool.

In this lesson you will

Build a simple routing rule that sends each request to the cheapest capable DeepSeek model
Identify the cost drivers in a DeepSeek bill and cut the biggest one
Decide when to replace another provider's subscription with DeepSeek, and when not to
Spot the red flags of overpaying and act on them

DeepSeek is cheap, which creates its own trap: bills small enough that nobody looks at them, running on habits that would be expensive elsewhere. This lesson is about getting the result you need at the lowest cost — including the cases where the lowest cost is not DeepSeek at all.

Figure 1.Four things drive an API bill: how many requests, how much context each carries, how much output each produces, and which model handles it. With DeepSeek, model choice (chat vs reasoner) and context size dominate.

Driver 1: the reasoner tax

On our catalogue prices at the time of writing, R1 0528 output costs about five times V3.2 output per token, and a reasoner generates several times more output tokens per answer because the thinking is billed. A request that costs a tenth of a cent on the chat model can cost a few cents on the reasoner. If half your traffic goes to the reasoner out of caution, you are paying many times what the work requires.

Figure 2.Output is where the money goes. The gap between Flash and R1 is more than ten to one before you count the extra reasoning tokens a reasoner produces.

Routing: cheap first, escalate on failure

Figure 3.Route every request to the cheapest model first. Only requests that fail a check — a parse error, a low confidence score, a human thumbs-down — go to the next tier.

The routing rule that works for most DeepSeek users has three tiers: V4 Flash for bulk, low-difficulty tasks; V3.2 or V4 Pro as the default for everything interactive; the reasoner only for requests that the default gets wrong or that are tagged as maths, logic or hard debugging. Implement the failure check as something mechanical — the output failed to parse as JSON, a validator rejected it, a user clicked retry. Do not route on vibes.

Tier 1 (Flash): classification, extraction, per-document summaries, rewriting to a template.
Tier 2 (V3.2 / V4 Pro): general chat, coding, drafting, combined analysis.
Tier 3 (reasoner): anything tier 2 failed twice, plus tasks explicitly tagged as reasoning-heavy.

Caching and off-peak batching

Lesson 7 covered both. As policy: put stable content first in every prompt so context caching bills it at the cache-hit rate; move any job that can wait into the off-peak window; and check whether your host offers a batch tier — note that in our catalogue, the batch rows for some third-party hosts are not cheaper than the standard rows, so read the ladder rather than assuming batch means discount.

Trimming context

Because V4 has a 1M-token window, it is tempting to send everything. Do not. Summarise long documents once with Flash and send the summary. Truncate chat history to the last few turns plus a running summary. Send the function, not the file. Every input token you cut is cut on every call for the life of the app.

When to move to DeepSeek — and when not to

If you pay a monthly subscription to another provider and use it mainly for text, code and analysis, try the free DeepSeek app for a week and cancel if you do not miss anything. If you run an API workload on a closed model for summarising, classification or extraction, benchmark the same prompts on V3.2 or V4 Flash: at catalogue prices the savings are often more than 90%, and our DeepSeek vs Claude cost and DeepSeek V4 vs GPT-5 cost pages show the arithmetic.

Do not move if you need strict data residency and cannot use a regional host, if your workload depends on image or voice features DeepSeek lacks, or if you have measured a quality gap on your own data that costs you more than the savings. Cheap output that has to be redone by a person is not cheap.

Tools for this lesson

AI cost calculator →Enter your request counts and token sizes to see a monthly DeepSeek bill by model.Compare models →Side-by-side prices and context windows across every provider we track.DeepSeek pricing page →Live catalogue prices, freshness status and pricing history for every DeepSeek row.DeepSeek vs GPT-4o cost →The workload-by-workload arithmetic for replacing a closed model with DeepSeek.Decision engine →Answer a few questions about your use and get a recommended plan and model.

Knowledge check

An API log shows that 60% of requests go to deepseek-reasoner and the average prompt is 25K tokens, mostly a fixed reference document. What is the first fix?

Lesson FAQ

Is DeepSeek cheaper than the OpenAI API?

On catalogue prices at the time of writing, DeepSeek's chat models cost a small fraction of OpenAI's flagship models per token. Whether it is cheaper for you depends on whether quality on your task holds up; benchmark before switching. See [DeepSeek vs GPT-4o cost](/compare/deepseek-vs-gpt4o-cost).

Should I use DeepSeek batch pricing?

Check the ladder first. In our catalogue some third-party batch rows are not cheaper than standard rows. The official API's off-peak window is the discount that reliably applies to deferred work.

How do I stop the reasoner from running up the bill?

Route to deepseek-chat by default, escalate to deepseek-reasoner only on a mechanical failure signal, set max_tokens and a timeout on every reasoner call, and watch the share of reasoner traffic in your logs.